Joram Soch, BCCN Berlin / GRK 1589 / Charite Berlin

Model Selection for General Linear Models in fMRI Data Analysis

Cognitive neuroscience is the scientific discipline that aims to uncover the neurobiological processes underlying mental states, cognitive functions and psychic phenomena. In the last 25 years, functional magnetic resonance imaging (fMRI) has become the most popular method to study relationships between the human brain and the human mind. Data obtained from fMRI experiments are usually analyzed with general linear models (GLMs) which relate mental states in a psychological experiment – more precisely: experimentally controlled cognitive operations – to neural activity measured via fMRI – more precisely: the observed hemodynamic signal. This allows for localization of cognitive functions such as perception, memory and consciousness in the human brain. While this approach has led to a wealth of new insights into the functional neuroanatomy of numerous regions in the human brain, GLMs for fMRI are also inherently user-dependent. As statistical models, they are based on the researcher’s assumptions about which processes influence the measured signal, at which point in time and to what extent. Typically, several models are tested and the final model selection is performed rather arbitrarily. This subjectiveness – and therefore lack of objectivity – in the data analysis process poses a threat to the scientific validity of GLM-based fMRI research. Statistical model selection provides a solution to this problem by grounding model choice in objective criteria such as model accuracy, model complexity and generalizability. In this work, we develop cross-validated Bayesian model selection (cvBMS), a method that allows to choose the best GLM for a given group-level fMRI data set. Further, we propose cross-validated Bayesian model averaging (cvBMA), a technique that allows to improve parameter estimates by combining several GLMs for a given subject-level fMRI data set. Finally, we describe software for model assessment, comparison and selection (MACS) that strengthens formal model quality control for GLMs applied to fMRI. Taken together, these methods will increase the reproducibility of cognitive neuroscience research.

Additional Information

PhD defence in the RTG 1589 "Sensory Computation in Neural Systems".

Organized by

John-Dylan Haynes / Robert Martin